Data Scientist Professional Practical Exam

Company Background

Inn the Neighborhood is an online platform that allows people to rent out their properties for short stays. At the moment, only 2% of people who come to the site interested in renting out their homes start to use it.

The product manager would like to increase this. They want to develop an application to help people estimate how much they could earn renting out their living space. They hope that this would make people more likely to sign up.

Customer Question

The product manager would like to know:

• Can you develop a way to predict how much someone could earn from renting their property that could power the application?

Success Criteria

They want to avoid estimating prices that are more than 25 dollars off of the actual price, as this may discourage people.

Dataset

The data you will use for this analysis can be accessed here: "data/rentals.csv"

Importing and cleaning the dataset

(8111, 9)

	id	latitude	longitude	property_type	room_type	bathrooms	bedrooms	minimum_nights	price
0	958	37.76931	-122.43386	Apartment	Entire home/apt	1.0	1.0	1	$170.00
1	3850	37.75402	-122.45805	House	Private room	1.0	1.0	1	$99.00
2	5858	37.74511	-122.42102	Apartment	Entire home/apt	1.0	2.0	30	$235.00
3	7918	37.76669	-122.45250	Apartment	Private room	4.0	1.0	32	$65.00
4	8142	37.76487	-122.45183	Apartment	Private room	4.0	1.0	32	$65.00

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 8111 entries, 0 to 8110
Data columns (total 9 columns):
 #   Column          Non-Null Count  Dtype  
---  ------          --------------  -----  
 0   id              8111 non-null   int64  
 1   latitude        8111 non-null   float64
 2   longitude       8111 non-null   float64
 3   property_type   8111 non-null   object 
 4   room_type       8111 non-null   object 
 5   bathrooms       8099 non-null   float64
 6   bedrooms        8107 non-null   float64
 7   minimum_nights  8111 non-null   int64  
 8   price           8111 non-null   object 
dtypes: float64(4), int64(2), object(3)
memory usage: 570.4+ KB

Checking null bathroom values

	id	latitude	longitude	property_type	room_type	bathrooms	bedrooms	minimum_nights	price
139	144978	37.79336	-122.42506	Apartment	Private room	NaN	1.0	30	$56.00
181	229240	37.79341	-122.40340	Hostel	Shared room	NaN	1.0	1	$45.00
196	259621	37.79470	-122.40374	Hostel	Shared room	NaN	1.0	1	$45.00
197	259622	37.79441	-122.40473	Hostel	Shared room	NaN	1.0	1	$45.00
267	430692	37.75906	-122.40761	Apartment	Private room	NaN	1.0	31	$68.00
352	596042	37.79384	-122.42436	Apartment	Private room	NaN	1.0	30	$63.00
434	785901	37.79313	-122.40443	Hostel	Shared room	NaN	1.0	1	$45.00
435	786492	37.79421	-122.40310	Hostel	Shared room	NaN	1.0	1	$45.00
436	786506	37.79260	-122.40339	Hostel	Shared room	NaN	1.0	1	$45.00
539	1031899	37.74833	-122.42621	Apartment	Entire home/apt	NaN	1.0	30	$127.00
606	1206233	37.77028	-122.44757	Apartment	Private room	NaN	1.0	1	$79.00
7036	34902361	37.77790	-122.43688	Apartment	Private room	NaN	1.0	30	$50.00

array([ 1. ,  4. ,  1.5,  2. ,  3. ,  0. ,  2.5,  3.5,  nan,  0.5,  6.5,
       10. ,  4.5, 14. ,  8. ,  5. ,  6. ,  7. ])

Checking null bedroom values

	id	latitude	longitude	property_type	room_type	bathrooms	bedrooms	minimum_nights	price
269	431862	37.78321	-122.41969	Apartment	Entire home/apt	1.0	NaN	30	$124.00
6301	32183178	37.78883	-122.48640	House	Entire home/apt	3.5	NaN	30	$650.00
7786	38329898	37.78347	-122.41669	Apartment	Entire home/apt	1.0	NaN	30	$75.00
7840	38550933	37.78979	-122.41994	Apartment	Entire home/apt	1.0	NaN	30	$108.00

array([ 1.,  2.,  0.,  3.,  4., nan,  5.,  6., 14.,  7.,  8.])

Checking other unique values

array(['Apartment', 'House', 'Condominium', 'Townhouse', 'Loft',
       'Guest suite', 'Cottage', 'Hostel', 'Guesthouse',
       'Serviced apartment', 'Bungalow', 'Bed and breakfast', 'Hotel',
       'Boutique hotel', 'Other', 'Tiny house', 'Resort', 'Villa',
       'Aparthotel', 'Castle', 'Camper/RV', 'In-law', 'Earth house',
       'Cabin', 'Dome house', 'Hut'], dtype=object)

array(['Entire home/apt', 'Private room', 'Shared room', 'Hotel room'],
      dtype=object)

Data validation

The dataset has 8111 rows and 9 columns, 16 missing values.

• id: Int, no null value.
• latitude: Float, no missing values.
• longitude: Float, no missing values.
• property_type: object, no missing values. Needs to change to categorical type.
• room_type: object, no null value. Needs to change to categorical type.
• bathrooms: float64, 12 missing values. Keep as float to count half baths and fill missing values with 0
• bedrooms: float64, 4 missing values. Remove missing values and change to int.
• minimum_nights: int, no missing values. Remove values above 60 days.
• price: object. replace "$" and "," and change to float. Remove prices = 0.

Checking minimum nights

array([        1,        30,        32,         6,         3,        90,
               2,         5,         4,        60,        10,       365,
              80,        45,         7,        29,        31,         9,
              14,       183,       200,       180,       120,        58,
             360,        50,        59,        70,        16,        75,
             110,        13,        55,       140,        28,      1125,
              85,        21,        18, 100000000,        11,       188,
            1000,        40,        65,       100,        12,        38,
               8,        25,        15,       150,        33], dtype=int64)

According to San Francisco administrative code, a short-term residential rental is a rental of all or a portion of your home for periods of less than 30 nights. For this reason, we will also create a new column to classify the data as short_term or not.

4528

Changing data types

count     4528.000000
mean       263.161661
std        479.444582
min          0.000000
25%        105.000000
50%        169.000000
75%        280.000000
max      10000.000000
Name: price, dtype: float64

Exploratory analysis

Overview

I have investigated the target variable and features of the properties, and the relationship between target variable and features. After the analysis,I decided to apply the following changes to enable modeling:

• price: use power transformation to approximate to a normal.
• Create a new ordinal variable from minimum_nights variable.
• Change latitude and longitude to be 0 centered and 10x bigger.

Target variable: Price

Text(0.5, 1.0, 'Normalized Rental Price Boxplot')

Minimum nights

array([0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. , 1.1, 1.2,
       1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2. , 2.1, 2.2, 2.3, 2.4, 2.5,
       2.6, 2.7, 2.8, 2.9, 3. , 3.1, 3.2, 3.3, 3.4, 3.5, 3.6, 3.7, 3.8,
       3.9, 4. , 4.1, 4.2, 4.3, 4.4, 4.5, 4.6, 4.7, 4.8, 4.9, 5. , 5.1,
       5.2, 5.3, 5.4, 5.5, 5.6, 5.7, 5.8, 5.9, 6. , 6.1, 6.2, 6.3, 6.4,
       6.5, 6.6, 6.7, 6.8, 6.9, 7. , 7.1, 7.2, 7.3, 7.4, 7.5, 7.6, 7.7,
       7.8, 7.9])

([<matplotlib.axis.YTick at 0x2f529da0250>,
  <matplotlib.axis.YTick at 0x2f529d97a90>,
  <matplotlib.axis.YTick at 0x2f529dcfe80>,
  <matplotlib.axis.YTick at 0x2f529dd8430>,
  <matplotlib.axis.YTick at 0x2f529dd8a30>,
  <matplotlib.axis.YTick at 0x2f529dd01c0>,
  <matplotlib.axis.YTick at 0x2f529dd0910>],
 [Text(0, 1.0, '$10'),
  Text(0, 1.5, ''),
  Text(0, 2.0, '$100'),
  Text(0, 2.5, ''),
  Text(0, 3.0, '$1000'),
  Text(0, 3.5, ''),
  Text(0, 4.0, '$10000')])

room_type	Entire home/apt	Hotel room	Private room	Shared room
property_type
Aparthotel	3	2	17	0
Apartment	667	0	424	4
Bed and breakfast	0	12	17	14
Boutique hotel	15	60	189	0
Bungalow	8	0	2	1
Cabin	1	0	1	0
Castle	0	0	4	0
Condominium	341	0	193	1
Cottage	7	0	2	0
Dome house	1	0	0	0
Earth house	0	0	1	0
Guest suite	421	0	64	0
Guesthouse	19	0	1	0
Hostel	0	35	9	23
Hotel	2	35	110	0
House	664	0	896	7
In-law	1	0	0	0
Loft	34	0	11	1
Other	1	0	17	0
Resort	0	2	11	0
Serviced apartment	49	18	4	0
Tiny house	2	0	0	0
Townhouse	41	0	56	0
Villa	1	0	5	0

(0.0, 60.0)

Latitude/longitude

Categorical variables: Property type, room type

Property types

Some property types has almost no samples, making it harder for the regression model to adjust using the property type.

I did some analysis on how the lack of samples did affect the predictions and created another column (adj_property_type) with less types.

Each category with less than 100 samples was mapped into the ones with the closest slope (MSE) compared to the nearby property types.

House                 1567
Apartment             1095
Condominium            535
Guest suite            485
Boutique hotel         264
Hotel                  147
Townhouse               97
Serviced apartment      71
Hostel                  67
Loft                    46
Bed and breakfast       43
Aparthotel              22
Guesthouse              20
Other                   18
Resort                  13
Bungalow                11
Cottage                  9
Villa                    6
Castle                   4
Cabin                    2
Tiny house               2
Earth house              1
Dome house               1
In-law                   1
Name: property_type, dtype: int64

Adjusting the property types

We will eliminate any property type with less than 100 samples and find the best fit (best MSE) to the linear regression of bedrooms x price_normed.

To check only the properties in the same region, we will only check within a margin from the tested dataset.

Serviced apartment --> Apartment
Hostel --> House
Loft --> Condominium
Bed and breakfast --> House
Aparthotel --> Townhouse
Guesthouse --> Condominium
Other --> Apartment
Bungalow --> House
Resort --> Hotel
Cottage --> Hotel
Villa --> Condominium
Castle --> House
Cabin --> Townhouse
Tiny house --> Condominium
Earth house --> House
Dome house --> Townhouse
In-law --> Boutique hotel

Room type

The most common room_type is the "Entire home/apt"

Relationship between property_type, room type and price.

Location x price EDA

Overview

array([[ 44.85661668],
       [ 65.91886454],
       [102.1101941 ],
       [169.54019535],
       [309.75507112],
       [651.19599368]])

I checked the location influence on price, and the median price of a single room close to the property has a strong influence on the final price.

To map the regions, we will divide the map in cuts of 0.04 latitude and longitude and check the median price near each point.

Region mapping

Useful functions

Single bedroom price mapping

For a visual understanding of the region mapping, we will make an example of the single bedroom median for our dataset.

As this mapping uses the whole dataset, we will have to re-create it after separating the train/test datasets.

Density mapping

Prepare Data for Modelling

To enable modelling, we chose id,latitude, longitude, adj_property_type, room_type, bathrooms, bedrooms, minimum_nights as features and price_normed as target variable. I also have made the following changes:

• Convert the categorical variables into numeric features
• Split the data into a training set and a test set
• Create region mappings using the training set
• Remove the price feature from X_train and X_test

Predicting the price is a regression problem in machine learning. I am choosing the Linear Regression model because we can see strong to moderate relationship between some features and target variable. The comparison model I am choosing is the Gradient Boosting Regressor model because it is easy to interpret and it generally provides better accuracy.

For the evaluation, I am choosing R squared and RMSE (Root Mean Squared Error) to evaluate the model. R squared measures how well the model fits dependent variables (i.e. features). RMSE measures how much your predicted results deviate from the actual number.

Generate the medium price regions using the train set

Checking property type 5
Checking property type 3
Checking property type 1
Checking property type 2
Checking property type 0
Checking property type 4
Checking property type 6
Done checking property types
Checking room type 0
Checking room type 2
Checking room type 1
Checking room type 3
Done checking room types

Base model - Linear regression

LinearRegression()

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Linear Regression r2_score:  0.56
Linear Regression Root Mean Squared Error:  0.65

Linear Regression predictions:
Within $25: 494
Total:      1359
Percentage: 36.4%

Decision tree regressor

Gradient Boost Regressor

array([[ 0.76120252],
       [-0.41541559],
       [-0.55874855],
       ...,
       [-0.4656158 ],
       [-0.18177027],
       [ 1.47861684]])

Best max_depth is: 5 
Testing R^2 is: 0.6136158440651533

GradientBoostingRegressor(max_depth=5, n_estimators=300, random_state=42)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Gradient Boosting r2_score:  0.61
Gradient Boosting Root Mean Squared Error:  0.61

Gradient Boosting predictions:
Within $25: 534
Total:      1359
Percentage: 39.3%

D:\Programs\python\lib\site-packages\sklearn\base.py:443: UserWarning: X has feature names, but PowerTransformer was fitted without feature names
  warnings.warn(

Text(0.5, 1.0, 'Gradient Boost Regressor Feature Importance')

The R squared of the Linear Regression, and the Gradient Boosting Regressor model is 0.55 and 0.61, meaning the Gradient Boosting Regressor model fits the features better.

The RMSE of the Linear Regression, and the Decision Tree Regression model is 0.65 and 0.61, meaning the Decision Tree Regression model has less error in predicting values.

Evaluate by Business Criteria

The company wants to avoid prediction out of a range - more than \$25 off from the actual price. Therefore, we would consider using percentage of predictions which predicted price is not more than \\$25 off from the actual price as a KPI to compare the two models again. The higher the percentage, the better the model performs. 39% of the decision tree regressor prediction is not more than $25 off from the actual rent price, while the linear regression model only have 36%.

Recommendations

To help the user predict the price, we can deploy this GradientBoostingRegressor model into production. By implementing this model, about 39% of the users will be encouraged to rent the property. to better evaluate wither this model can really encourage more people to rent their property, I would also recommend A/B testing about using this model to compare two groups of users.

To implement and improve the model, I will consider the following steps:

The ideal way to deploy this model is to deploy as a web services in the home page of the website, so we can see the effects of the model on having new users renting properties on the platform. The calculator should lead to the registering page, auto-completing the fields filled before to make it as friction-less as possible.

Collecting more data, e.g. area, property age, electronics, internet, garage, demand and seasonality has a big impact on the rental prices for short stays. Some categories had just a single sample. Ideally, having several examples for each possible room_type x property_type combination would improve the predictions.

Feature Engineering, e.g reduce the categories in model, have several region mappings for each category to have a "region classification".

Improve the calculator by showing a price range according to the region variance instead of a single prediction, so the user can also have an idea of the price ranges in the region and use his own knowledge of intangible values such as brand, design and maintenance to set the price.

room_type	Entire home/apt	Hotel room	Private room	Shared room
adj_property_type
Apartment	214.0	5.0	142.0	1.0
Boutique hotel	3.0	13.0	69.0	NaN
Condominium	110.0	NaN	80.0	2.0
Guest suite	117.0	NaN	20.0	NaN
Hotel	3.0	17.0	38.0	NaN
House	181.0	4.0	283.0	14.0
Townhouse	12.0	NaN	31.0	NaN

Using TPOT

D:\Programs\python\lib\site-packages\sklearn\utils\validation.py:1111: DataConversionWarning: A column-vector y was passed when a 1d array was expected. Please change the shape of y to (n_samples, ), for example using ravel().
  y = column_or_1d(y, warn=True)

Optimization Progress:   0%|          | 0/210 [00:00<?, ?pipeline/s]

Generation 1 - Current best internal CV score: 0.8003962038422999

Generation 2 - Current best internal CV score: 0.801019490879051

Generation 3 - Current best internal CV score: 0.801019490879051

Generation 4 - Current best internal CV score: 0.801019490879051

Generation 5 - Current best internal CV score: 0.8026510607366657

Generation 6 - Current best internal CV score: 0.8027330906678115

Generation 7 - Current best internal CV score: 0.8027330906678115

Generation 8 - Current best internal CV score: 0.8027330906678115

Generation 9 - Current best internal CV score: 0.8027330906678115

Generation 10 - Current best internal CV score: 0.8027330906678115

Best pipeline: RandomForestRegressor(KNeighborsRegressor(LinearSVR(input_matrix, C=1.0, dual=True, epsilon=0.001, loss=squared_epsilon_insensitive, tol=0.01), n_neighbors=18, p=2, weights=uniform), bootstrap=False, max_features=0.3, min_samples_leaf=3, min_samples_split=15, n_estimators=100)

TPOTRegressor(generations=10, offspring_size=20, population_size=10,
              scoring='r2', verbosity=2)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Linear Regression r2_score:  0.62
Linear Regression Root Mean Squared Error:  0.6

Linear Regression predictions:
Within $25: 533
Total:      1359
Percentage: 39.2%

D:\Programs\python\lib\site-packages\sklearn\base.py:450: UserWarning: X does not have valid feature names, but LinearSVR was fitted with feature names
  warnings.warn(
D:\Programs\python\lib\site-packages\sklearn\base.py:443: UserWarning: X has feature names, but PowerTransformer was fitted without feature names
  warnings.warn(

Using Hyperopt

100%|██████████| 100/100 [05:14<00:00,  3.15s/trial, best loss: 0.58719646799117] 

{'max_depth': 10.0, 'min_samples_split': 43, 'n_estimators': 330.0}

Linear Regression r2_score:  0.58
Linear Regression Root Mean Squared Error:  0.63

Linear Regression predictions:
Within $25: 548
Total:      1359
Percentage: 40.3%

D:\Programs\python\lib\site-packages\sklearn\base.py:443: UserWarning: X has feature names, but PowerTransformer was fitted without feature names
  warnings.warn(